The current issue and full text archive of this journal is available on Emerald Insight at:

www.emeraldinsight.com/1526-5943.htm

JRF
17,2
Credit risk signals in CDS
market vs agency ratings
Michael Jacobs Jr
Financial Advisory, Risk Models, Methodologies & Analytics,
194 Accenture Consulting, New York, New York, USA
Received 31 July 2015 Ahmet K. Karagozoglu
Revised 17 November 2015
4 January 2016
Department of Finance, Zarb School of Business, Hofstra University,
Accepted 17 January 2016 Hempstead, New York, USA, and
Dina Naples Layish
Department of Finance, School of Management, Binghamton University,
Binghamton, New York, USA

Abstract
Purpose – This research aims to model the relationship between the credit risk signals in the credit
default swap (CDS) market and agency credit ratings, and determines the factors that help explain the
variation in such signals.
Design/methodology/approach – A comprehensive analysis of the differences in the relative credit
risk assessments of CDS-based risk signals and agency ratings is provided. It is shown that the
divergence between credit risk signals in the CDS market and agency ratings is explained by factors
which the rating agencies may consider differently than credit market participants.
Findings – The results suggest that agency credit ratings of relative riskiness of a reference entity do
not always correspond with assessments by CDS spreads, as the price of risk is a function of additional
macro and micro factors that can be explained using statistical analysis.
Originality/value – This research is unique in modeling the relationship between the credit risk
assessments of the CDS market and the agency ratings, which to the best of the authors’ knowledge has
not been analyzed before in terms of their agreement and the level of discrepancy between them. This
model can be used by investors in debt instruments that are not explicitly CDSs or which have illiquid
CDS contracts, to replicate market-based, point-in-time credit risk signals. Based on both market-based
and firm-specific factors in this model, the results can be used to augment through-the-cycle credit risk
assessments, analyze issues surrounding the pricing of CDSs and examine the policies of credit rating
agencies.
Keywords CDS, Credit rating agencies, Credit risk, Credit default swap, Credit ratings
Paper type Research paper

1. Introduction
The primary purpose of a credit default swap (CDS) contract is to provide protection to
the purchaser of a debt instrument in case of default or a related credit event, serving as

The Journal of Risk Finance JEL classification – G10, G20, G24, G28, G32
Vol. 17 No. 2, 2016
pp. 194-217 The views expressed herein are those of the authors and do not necessarily represent a position
© Emerald Group Publishing Limited
1526-5943
taken neither by Accenture Consulting, nor of any affiliated firms. Karagozoglu acknowledges
DOI 10.1108/JRF-07-2015-0070 Zarb School of Business Summer Research Grant, which partially supported this article.
a form of insurance. A CDS contract can also be used on the short side to bet against the Credit risk
credit quality, or to hedge a long position in the debt or equity of a reference entity. An signals
investor in a CDS contract pays an annual premium to the seller of the contract. If a
credit event such as default of the underlying reference entity occurs, the seller buys the
underlying debt instrument from the investor at par. The annual premium, or CDS
spread, ultimately reflects the market price of the credit risk with respect to the
underlying instrument. 195
Credit risk makes up perhaps the largest risk an investor bears when buying a
defaultable fixed-income instrument. Credit risk may be broadly defined as the
uncertainty associated with potential loss of value on a fixed-income obligation – either
principle or interest – in the event of default, downgrade or widening of credit spreads.
Prior to the beginning of CDS trading, ratings assigned by agencies were the only
signal of the credit risk of fixed-income instruments. Credit ratings are, in theory, an
independent assessment of the relative credit risk of a firm. There is a natural
expectation that the CDS spread on a specific debt instrument will be correlated to the
credit rating of the underlying reference entity. The primary difference lies in the
timeliness of the information: credit ratings are updated periodically, whereas CDS
spreads are continuously updated through ongoing trading, providing a current
measure of the market’s interpretation of the risk of the debt instruments.
According to the policy and guidelines issued by the nationally recognized statistical
rating organizations (NRSROs)[1], at any given time, the credit rating on an issue of debt
reflects its relative credit quality over some horizon. This has the interpretation that a
credit rating embodies information on the obligor’s probability of default (PD) relative to
a cohort, potentially allowing for a standard comparison of default risks. Therefore,
ratings represent an opinion regarding potential loss, a firm’s capacity to pay back all its
sources of financing as well as the recovery of a particular instrument in the event of
default (Micu et al., 2006). Historically, Standard & Poor’s (S&P) has primarily issued a
senior unsecured debt rating, presumably a ranking of pure default risk; in cases where
there is subordinated debt, a separate rating is issued that may be lower than that issued
to the senior unsecured debt, to reflect the greater recovery risk.
The agencies claim that they modify a firm’s relative credit rating only if a change
occurs in a borrower’s fundamental creditworthiness, implying that they do not react to
systematic events, which affect all firms equally but do not impact relative credit
quality[2]. In addition to issuing credit ratings, rating agencies issue rating reviews and
outlooks; these announcements follow the occurrence of material events that potentially
could have an impact on a firm’s fundamental credit quality and signal a possible rating
change. Furthermore, rating agencies base their rating assignments on many different
factors, some public, as in financial statements or capital markets information, and some
private, as in an assessment of management quality or industry position (Hull et al.,
2004).
Following the subprime debacle starting in the summer of 2007, rating agencies have
come under much scrutiny mainly because the market began to question the validity of
the ratings that were issued (Hull, 2009; Stephens, 2012). A key regulation relevant to
this research is The Dodd–Frank Wall Street Reform and Consumer Protection Act of
2010 (Dodd-Frank, 2010), perhaps the most ambitious and far-reaching overhaul of
financial regulation since the 1930s. The main purposes of this legislation include
identifying and regulating systemic risk through a special council that can deem
JRF non-bank financial firms as systemically important, regulate them and as a last resort
17,2 break them up. Furthermore and salient to this study, Dodd–Frank imposes a new
regulatory scheme on rating agencies and tightens existing regulation, with the primary
goals of holding rating agencies accountable for the quality of their credit ratings and
enhancing the transparency of credit ratings.
More recently, in February of 2013, the US Government sued S&P Ratings Services
196 over the quality of the actual ratings that were issued during the financial crisis
(Eaglesham et al., 2013). Much of the scrutiny on the rating agencies centers on their
inability to properly rate structured mortgage and commercial credit, such as
collateralized mortgage obligations, collateralized debt and collateralized loan
obligations. One would expect such a deficiency to also be reflected in the single-name
corporate credit market.
We note the significance of the Basel III supervisory guidance (BCBS, 2010) for our
study of CDS credit signals and agency credit ratings. A number of measures mitigate
the reliance on external ratings in the Basel II (BCBS, 2006) framework. The measures
include requirements for banks to perform their own internal assessments of externally
rated exposures, the elimination of certain “cliff effects” associated with credit risk
mitigation practices and the incorporation of key elements of the International
Organization of Securities Commissions (IOSCO) code of conduct fundamentals for
credit rating agencies (IOSCO (2004)) into the eligibility criteria for the use of external
ratings in the capital framework. Included in this set of measures are market-related
monitoring tools, such as CDS spreads, which provide a source of instantaneous data on
potential liquidity difficulties, useful data to monitor asset prices and liquidity. It is
evident that international regulators recognize not only that CDS spreads are a useful
complement to risk ratings, but also that CDS spreads may provide qualitatively
different information, such as liquidity or other market risk–related dimensions that
could influence default risk.
Understanding the relationship between agency credit ratings and CDS spreads
credit risk signals can help explain how market participants perceive and price credit
risk. Considerable research has analyzed the relationship. As noted by Callen et al.
(2009), although CDS spreads are related to credit ratings issued by rating agencies,
among firms having a given rating, there is quite a wide variation in CDS spreads. Cizel
(2013) indicates that the CDS spreads bear the closest correspondence to the market
assessment of firms’ credit risk. Therefore, if CDS spreads reflect a component of pure
credit risk (i.e. the risk of loss associated with a deterioration in credit quality or default
on the reference entity’s debt), and credit ratings quantify the relative likelihood of a
corporation defaulting on its debt, then the CDS contracts of reference entities with a
given credit rating should be priced similarly. The research in this paper models the
relationship between the credit risk signals in the swap market and the agency ratings
and determines the factors that help explain the variation in such signals.
Hilscher and Wilson (2013) suggest that agency ratings do not clearly separate firms
into categories by their PD, especially for the investment-grade issuers. Their
conclusion that any single measure cannot accurately reflect all relevant aspects of
credit risk provides a strong motivation to analyze the credit risk measures from the
swap market and the rating agencies. This paper contributes to the literature by
providing empirical results for the relationship between credit risk signals obtained
from two measures, CDS-based ratings versus agency ratings risk categories.
 There are several studies in the literature that have looked at the relationship Credit risk
between credit spreads, bond yields and agency ratings. Hull et al. (2004) examine the signals
theoretical relationship between bond yields and CDS spreads, including how this is
influenced by rating agency announcements. The authors find evidence that the CDS
market anticipates all three types of negative credit events – downgrade, negative watch
and negative outlook – on the announcement day[3]. Daniels and Shin-Jensen (2005)
study the relationship between CDS spreads, credit spreads of corporate bonds and 197
credit rating changes. They find that downgrades significantly impact spreads, and that
this effect is accentuated for investment-grade issues. Using the Treasury yield curve as
a proxy for the risk-free term structure, they also illustrate the dependence of the value
of the CDS contract on the risk-free rate[4].
Generally, prior studies identified several of the most common variables found to
affect CDS spreads: the leverage of the reference entity and option-implied volatility of
its equity, the risk-free rate and liquidity of the CDS contract. Villouta (2006)
investigates the pricing effects of liquidity in the corporate bond and CDS markets and
finds that highly liquid CDS contracts tended to have lower CDS bases, and illiquid
contracts higher bases. Carr and Wu (2010) find that equity option-implied volatility and
CDS spreads covary positively, with credit markets sometimes showing variation
independent of the stock and option markets. Cao et al. (2010) argue that CDS contracts
are similar to out-of-the-money put options and find that put option-implied volatility
dominates historical volatility in explaining the time-series variation in CDS spreads.
Das and Hanouna (2008) find that variables most correlated with recovery rates (i.e.
losses given default, LGDs) are the one-month risk-free rate, the CBOE volatility index
(VIX), the yield curve and correlation between the levels in the term structure and the
equity market volatility.
Schneider et al. (2007) examine the relationship between LGD and PD as implied by
CDS spreads, and use agency ratings as a proxy for PD (i.e. as a crude representation for
credit quality). They find evidence that equity market volatility, as measured by the
VIX, is positively correlated to long- and short-term default factors that directly
influence the valuation of CDS.
A CDS contract allows an investor to trade solely on the credit risk of a firm,
especially as an investor does not need to hold the underlying debt contract to trade in
the CDS market. This fact has led many to conclude that the overall risk in the CDS
market is heightened: there is less of an incentive to monitor borrowers when insurance
can easily be acquired (Stulz, 2010), and many investors can purchase a CDS contract on
the same underlying asset. Mahfoudhi (2011) examines new market conventions in the
CDS market, the so-called CDS “big bang” changes to contracts, meant to reinforce
confidence and ensure its long-term growth. Benzschawel and Corlu (2011) also examine
this phenomenon, pointing out that conventions for trading CDS have enabled investors
to go short with little or no initial investment, contributing to the unprecedented
volatility in cash and synthetic credit markets since mid-2007.
Resti and Sironi (2007) indicate that agency ratings are less reactive than CDS
spreads to changes in credit risk levels. These facts lead us to hypothesize that
variations in CDS-based risk signals for a given agency rating may exist, and a
comprehensive analysis of these variations would be useful for both market participants
and regulators. Furthermore, the divergence between CDS spreads and agency credit
ratings are apt to be explained by factors viewed differently by the rating agencies and
JRF credit market participants. Under certain ideal conditions, theoretically, the CDS spread
17,2 should represent pure credit risk.
Understanding the relationship between CDS spreads and agency credit ratings can
help explain how market participants perceive and price credit risk. Our research
examines possible contributing factors that influence this range. While modeling and
analyzing the relationship between the credit risk signals in CDS spreads and agency
198 ratings, this study:
• incorporates an overview of different methodologies used to value CDSs;
• addresses the variables considered to influence the value of the CDS; and
• connects the influential variables to the explanation of the range that exists
between CDS spreads of reference entities with the same ratings.

This paper analyzes differences in the relative credit risk assessments, identified as the
agency credit rating categories (ACR-c) versus the CDS categories (CDS-c). The analysis
suggests that the market perceives the average CDS in a given ACR-c as riskier than its
agency credit rating would dictate. These observations were evident throughout our
data period from February 28, 2003, to December 31, 2010. Two models were used to
explain the factors that could lead to this observation, having variables that would
either be used to explain the ACR-c being equal to the CDS-c, or the CDS-c being a more
severe assessment of credit risk than the ACR-c. Results for both of the models are
statistically and economically significant, and had signs on (magnitudes of) coefficient
estimates that are economically intuitive (significant). The models yield the following
results:
• The VIX, term premium or the put option-implied volatility of the associated
equity decreases the chances (predicted number of notches) of agreement, and
increases relative pessimism on the part of the CDS market.
• Conversely, the level of the S&P 500 Index, market capitalization or the earnings
per share (EPS) estimate of the reference entity, as well as the seniority of the debt,
are always negatively related to the former target variables.
• The crisis and post-crisis dummy variables show that agreement decreases
between the credit risk signals from the CDS market and agency ratings during
the financial crisis, and risk signals appear to be getting closer to each other after
market turbulence. When considered with respect to the periods between agency
rating changes, results show that the level of agreement decreases and distance
between the credit risk categories increases with the length of the duration.

We can highlight a potential use of our model as another element for practitioners in
their modeling tool-kit. This element would be especially valuable in the context of
portfolios of credit instruments only internally rated, where many practitioners will
have access to only through-the-cycle (TTC) credit ratings that may not necessarily
incorporate credit market data. Our model bridges the gap between TTC ratings such as
what the rating agencies produce, and so-called point-in-time (PIT) ratings. Therefore, if
these evaluations can be mapped to the rating agency grades, then the model can be used
to augment such credit assessments, thereby producing quasi-PIT ratings that can be
used in a variety of contexts such as pricing or trading credit instruments.
 The paper is organized as follows. Section 2 discusses our data and variable Credit risk
definitions. Section 3 presents the results of our empirical analysis, and Section 4 signals
concludes the paper.

2. Data and variable definitions

The data used in our study are obtained from Bloomberg. Our CDS sample consists of
contracts denominated in US dollars with a five-year term on reference entities having at 199
least one senior or subordinate issue, in addition to having US equity listings[5]. We
exclude contracts with less than 100 days of spread data. Our final data set contains the
daily five-year CDS spreads on 1,334 contracts on 392 distinct reference firms over
the period from February 28, 2003, to December 31, 2010. We end our sample just
after the resolution of the market turmoil in December 2010, as we are interested in
analyzing the discrepancies in credit risk signals in CDS market and agency ratings
through the global financial crisis (Ross, 2008), and also to avoid a period of new market
conventions in the CDS market (the “big bang”) in the years after the financial crisis
(Mahfoudhi 2011).
Additionally, we collect from Bloomberg the following firm-specific variables and
various market indicators:
• following Cizel (2013), the S&P long-term credit rating changes during our data
period for all reference entities[6];
• put option-implied volatility (POIV), which is consistent with previous research
that suggests equity market and option market volatilities are key components of
CDS valuation, for example Carr and Wu (2010) and Cao et al. (2010);
• market capitalization (MCAP);
• leverage (LEV), which is measured by the debt-to-equity ratio; and
• earnings per share estimate (EPSE)[7].

Our use of equity markets liquidity measure MCAP is motivated by Das and Hanouna
(2009), while use of the LEV and EPSE follows from the Merton (1974) basic structural
modeling framework for credit risk (i.e. increased LEV and decreased EPSE imply that
firms are nearer to their default points and hence indicate greater credit risk). We use a
reference-entity (issue-specific) variable to indicate the seniority of the debt (SDI). This
is motivated by Das and Hanouna (2008), who document the existence of a recovery or
LGD component of CDS spreads, as well as by Jacobs and Karagozoglu (2011), who
show that the LGD of corporate debt decreases with the seniority debt:
• the S&P 500 Index (SP500);
• the CBOE volatility index (VIX);
• the five-year swap yield premium (SYP), which is the difference between the
five-year plain vanilla swap rate and the five-year USA Treasury Note rate; and
• the term premium (TP), which represents the slope of the yield curve and is the
difference between the rates on five-year USA Treasury Note and three-month
USA Treasury Bill.

Zhang et al. (2009) demonstrate the significance of the S&P 500 Index in explaining CDS
spreads as part of a structural model, in which the index proxies for the unobservable
JRF state of the economy, with a higher index level implying lower level of systematic risk.
17,2 Schneider et al. (2007) find a strong positive relationship between CDS premia and levels
of the VIX. Hull et al. (2004) suggest that in CDS pricing, the optimal risk-free curve to
use lies between the Treasury and the swap curves; thus, we incorporate both measures
in our analysis. The rationale for using the TP variable comes from the fact that the
slope of the yield curve accounts for the risk premia due to investors’ time preference,
200 and also for their aversion to interest rate risk, which are expected to enter the pricing
kernels of most pricing models applicable to these assets, for example as in Cox et al.
(1981). Following Liu et al. (2006), the SYP is expected to account for the counterparty
risk, which is present between the CDS dealers and which is reflected in the higher
borrowing rates than investment counterparties pay, such as clearinghouses or banks,
as compared to the US Government. An increase in the SYP, all else equal, is expected to
augment CDS premia.
We create two dummy variables to investigate the CDS and rating relationship
during and after the financial crisis. The crisis dummy variable (CRIS) takes the value 1
for the period between June 2007 and June 2009, and the post-crisis dummy variable
(POST) takes the value 1 for the period between July 2009 and December 2010[8].
For us to compare credit risk signals from CDS spreads and agency ratings
(specifically S&P-assigned long-term ratings), we rescale the ratings to have five agency
credit rating categories (ACR-5c)[9]:

兵
1 ……………. AAA to A⫹
2 ……………. A to A⫺
ACR-5c ⫽ 3 for ratings BBB⫹ to BBB (1)
4 ……………. BBB⫺ to BB⫹
5 ……………. BB to CCC⫺

Both Huang et al. (2012) and Cizel (2013), like others, create CDS categories using the
agency-assigned ratings and treat those swaps within each category the same in terms
of their risk assessment. However, we show that spreads on swaps with the same rating
differ significantly, and therefore, our goal is to model these observed differences.
Hence, we create a 5-by-5 matrix of independent categories[10]. Based on the quintiles
of spread levels, we create five CDS categories (CDS-5c), where CDS-5c ⫽ 1 is assigned
to swap contracts with spread levels within the lowest 20th percentile on a given day,
and CDS-5c ⫽ 5 the highest (similar to Das and Hanouna, 2008). As a result of this
mapping, we are able to assign ACR-5c and CDS-5c to each contract for each trading day
for which we have the data. We create categories such that both ACR-5c ⫽ 1 (top
alphanumeric scale ratings) and CDS-5c ⫽ 1 (lowest quintile of spread levels)
correspond to the best credit risk signal (i.e. lowest risk). Similarly, ACR-5c ⫽ 5 (bottom
alphanumeric scale ratings) and CDS-5c ⫽ 5 (highest quintile of spread levels)
correspond to the worst credit risk signal (i.e. highest risk). If the CDS market’s and
rating agency’s assessments of the underlying reference entity’s credit quality are
similar, on average CDS-5c and ACR-5c should be the same.
To check the robustness of the results, and to ensure that they are not biased as a
result of our calibration of ACR categories, we also use deciles to create CDS-10c and
ACR-10c variables. The ACR-10c variables are defined as rescaled S&P long-term
credit ratings (i.e. a letter or alphanumeric scale) based on deciles of the ratings
distribution in our data set. Analogously, the CDS-10c variables are based on deciles of Credit risk
the spread level distribution, where CDS-10c ⫽ 1 is assigned to swap contracts with signals
spread levels within the lowest 10th percentile on a given day, and CDS-10c ⫽10 the
highest. Therefore, for the purpose of robustness checking, we create a 10-by-10 matrix
of independent categories.
Tables I and II present the credit risk signals cross-tabulated: Table I by ACR-5c and
CDS-5c, and Table II by ACR-10c and CDS-10c. If the CDS market’s and rating agency’s 201
assessments of the underlying reference entity’s credit quality are similar, on average,
CDS-5c and ACR-5c should be the same. If that is the case, most observations would be
concentrated on the diagonal cells of Tables I and II. Our sample contains 394,277
reference entity-days (approximately eight years of daily data on 1,334 swap contracts
with distinct underlying reference entities), and it can be seen that indeed the diagonals
are heavily populated relative to the off-diagonals. For example, in Table I for CDS-5c
versus ACR-5-c, only 8.1 per cent of the sample (32,049 of 394,277 reference entity-days)
has credit risk signals in both CDS market and agency ratings that suggest the best
credit quality (lowest risk), that is CDS-5c ⫽ 1 and ACR-5c ⫽ 1. Similarly, in Table II for

ACR-5c
Contract days 1 2 3 4 5 (%)

CDS-5c
1 32,049 23,698 20,122 4,172 2,295 82,336 20.88
2 19,387 34,818 22,939 5,823 2,790 85,757 21.75
3 13,411 20,239 34,630 8,632 3,254 80,166 20.33
4 9,247 13,176 18,970 29,884 7,298 78,575 19.93
5 2,954 3,782 4,091 23,504 32,112 66,443 16.85
77,048 95,713 100,752 72,015 47,749 394,277
(%) 19.54% 24.28% 25.55% 18.27% 12.11%

Notes: February 28, 2003, to December 31, 2010; agency credit ratings categories (ACR-5c) are defined as
rescaled Standard & Poor’s long-term credit (i.e., alphanumeric scale) ratings, given by equation (1):

兵
1…………….AAA to A⫹
2…………….A to A⫺
ACR-5c ⫽ 3 for ratings BB⫹ to BBB
4…………….BBB⫺ to BB⫹
5…………….BB to CCC ⫺

Credit default swap categories (CDS-5c) are based on the quintiles of spread levels, where CDS-5c ⫽ 1
is assigned to swap contracts with spread levels within the lowest 20th percentile on a given day, and
CDS-5c ⫽ 5 the highest 20th percentile; as a result of this mapping, we are able to assign ACR-5c and Table I.
CDS-5c to each contract for each trading day for which we have the data; we create categories such that Categories of credit
both ACR-5c ⫽ 1 (top letter ratings) and CDS-5c ⫽ 1 (lowest quintile of spread levels) correspond to the risk signals. Credit
best credit risk signal (i.e., lowest risk); similarly, ACR-5c ⫽ 5 (bottom letter ratings) and CDS-5c ⫽ 5 default swap (CDS-c)
(highest quintile of spread levels) correspond to the worst credit risk signal (i.e., highest risk); if the CDS vs agency credit
market’s and rating agency’s assessments of the underlying reference entity’s credit quality are similar, ratings (ACR-c):
on average, CDS-5c and ACR-5c should be the same; if that is the case, most observations would be CDS-5c based on
contained in the diagonal cells of this table quintiles vs ACR-5c
 JRF
17,2

202

Table II.

ratings (ACR-c):

based on deciles
vs agency credit
risk signals. Credit
Categories of credit

default swap (CDS-c)

CDS-10c vs ACR-10c
ACR-10c
Contract days 1 2 3 4 5 6 7 8 9 10 (%)

CDS-10c
1 13,457 5,051 4,018 3,305 2,818 1,316 143 219 2 0 30,329 7.69
2 6,402 13,594 5,688 4,903 1,731 1,500 331 331 42 0 34,522 8.76
3 2,834 4,343 14,374 5,083 3,526 2,015 999 449 216 41 33,880 8.59
4 2,891 5,811 5,694 7,889 7,304 5,510 3,680 2,069 660 512 42,020 10.66
5 2,951 6,721 6,419 7,495 9,426 5,701 4,303 3,102 1,134 474 47,726 12.10
6 2,708 4,380 5,775 5,993 9,453 12,965 4025 2729 1101 237 49,366 12.52
7 2,334 2,224 4,633 4,199 8,382 10,599 9,291 3,754 1,145 96 46,657 11.83
8 850 1,494 2,839 3,402 5,035 7,686 7,135 8,582 2,506 336 39,865 10.11
9 165 416 967 1,836 2,463 8,814 5,281 7,997 7,396 1,950 37,285 9.46
10 48 374 420 781 1,123 2,734 4,812 6,935 12,170 3,230 32,627 8.28
34,640 44,408 50,827 44,886 51,261 58,840 40,000 36,167 26,372 6,876 394,277
(%) 8.79% 11.26% 12.89% 11.38% 13.00% 14.92% 10.15% 9.17% 6.69% 1.74%

Notes: Agency credit ratings categories (ACR-10c) are defined as rescaled Standard and Poor’s long-term credit (i.e., alphanumeric scale) ratings based on the
deciles of distribution of ratings in our data set; credit default swap categories (CDS-10c) are based on the deciles of spread levels, where CDS-10c ⫽ 1 is assigned to
swap contracts with spread levels within the lowest 10th percentile on a given day, and CDS-10c ⫽ 10 to the highest 10th percentile
CDS-10c versus ACR-10-c, in 3.4 per cent of the sample (13,457 of 394,277 reference Credit risk
entity-days), credit risk signals in both sources suggest the best credit quality, that is signals
CDS-10c ⫽ 1 and ACR-10c ⫽ 1. Overall and by both measures, for the worst credit
quality categories, data appear to concentrate below the diagonal; that is, the CDS
market’s assessment is more severe, and in the case of the better credit quality
categories, the rating agency assessment appears to be more severe. A possible
explanation for this phenomenon is that the rating agencies may not downgrade these 203
counterparties as fast as the increases in CDS spreads, as their credit quality worsens,
and vice versa for improvements in credit quality.
The analysis of Tables I and II shows a test of robustness: the sample is restricted to
the subset where we observe an S&P initial rating or a rating change. Table III shows
analysis of CDS credit risk signals during two-month periods following either initial or
changed S&P ratings and tabulate what percentage of CDS-5c and ACR-5c provide the
same credit risk level. For example, in Table III for CDS-5c, 87.9 per cent of the sample
has credit risk signals in both CDS market and agency ratings that suggest the best
credit quality (lowest risk), that is CDS-5c ⫽ 1 and ACR-5c ⫽ 1. This high rate of
matching holds across categories. Therefore, we observe that in this restricted sample,
the credit risk categories as implied by agency credit ratings and CDS spreads match to
a significantly higher degree[11]. This finding supports our hypothesis that the factors
used in the model can be used to derive a CDS-like PIT credit risk signal, as this is
evidence that the level of agreement between credit ratings and CDS spreads is
significantly higher for this restricted sample of days following the agency revisions of
ratings. This observation suggests that the factors used in our models to derive PIT
credit risk signals might be used more explicitly by the rating agencies around their
rating decisions.
Table IV presents the summary statistics in basis points for the daily median swap
spreads for each of the CDS and agency credit ratings categories: Panel A for ACR-5c
and CDS-5c, and Panel B for ACR-10c and CDS-10c. As expected, the median swap
spreads increase with the worsening credit quality categories in both the CDS market

(%) ACR-5c
CDS-5c 1 2 3 4 5

1 87.9 1.5 0.0 0.0 0.0

2 6.7 90.3 2.9 0.0 0.0
3 5.4 6.2 87.2 1.9 1.0
4 0.0 2.0 8.5 84.3 17.0
5 0.0 0.0 1.0 13.8 82.0
100% 100% 100% 100% 100%
Table III.
Notes: February 28, 2003, to December 31, 2010; in this table, we consider the credit risk signals from Relationship between
the CDS market during the two-month period following S&P rating changes and tabulate what credit risk signals.
percentage of CDS-5c and ACR-5c provide the same credit risk level; we observe that credit risk Following rating
categories from agency credit ratings and CDS spreads are matching significantly higher for this changes. Credit
restricted sample; this finding supports our hypothesis that the factors used in the model can be used to default swap
derive CDS-like point-in-time credit risk signal holds; the level of agreement between agency credit (CDS-5c) vs agency
ratings and CDS spreads should be significantly higher for this restricted sample of days following credit ratings
agency revisions of ratings (ACR-5c)
JRF Minimum Mean Median Maximum SD Kurtosis Skewness
17,2
Panel A: CDS-5c based on quintiles vs ACR-5c
CDS categories
1 4.83 28.91 23.34 180.00 16.92 7.86 1.90
2 14.46 50.79 40.00 357.50 32.46 6.83 1.74
204 3 15.75 73.75 41.83 568.64 65.96 7.05 1.92
4 19.18 118.02 58.75 1048.43 126.86 9.97 2.46
5 36.39 388.91 220.00 9183.38 542.96 49.58 5.39
ACR categories
1 4.83 52.31 30.00 3349.11 93.94 52.27 11.73
2 7.84 65.70 34.00 9003.03 174.24 64.75 22.74
3 8.37 73.87 46.75 3443.68 106.25 94.25 10.40
4 15.51 160.32 105.00 6407.20 229.95 45.36 9.08
5 15.51 463.39 320.00 9183.38 591.30 39.80 4.83
Panel B: CDS-10c based on deciles vs ACR-10c
CDS categories
1 4.83 25.50 20.44 161.44 15.51 5.74 1.50
2 10.33 37.39 29.87 180.00 21.21 4.99 1.32
3 14.34 52.14 46.00 240.00 29.29 4.36 1.10
4 15.51 45.06 25.67 357.50 37.70 6.36 1.85
5 15.74 53.59 32.24 420.11 49.31 8.21 2.21
6 15.87 69.52 40.00 568.64 68.16 8.51 2.30
7 18.53 94.33 48.50 602.46 94.98 8.20 2.24
8 22.43 131.51 62.50 1048.43 144.28 8.93 2.37
9 36.30 212.40 103.18 1636.97 219.24 8.08 2.14
10 26.81 545.69 331.67 9183.38 685.83 33.02 4.47
ACR categories
1 4.83 49.11 30.45 2413.54 81.46 21.00 8.11
2 6.16 53.49 29.75 3349.11 98.10 68.49 12.36
3 7.84 50.41 31.00 1120.38 74.83 56.74 6.39
4 8.90 82.56 38.41 9003.03 239.03 57.86 18.12
5 8.37 60.29 40.67 1543.82 72.86 79.35 6.71
6 9.10 85.83 53.09 3443.68 127.48 65.48 10.03
7 15.51 130.01 87.75 2531.19 146.15 36.90 4.45
8 15.51 238.29 153.94 6407.20 378.77 81.47 7.59
9 15.51 515.38 368.35 9183.38 600.63 37.53 4.65
Table IV. 10 15.92 651.40 522.50 8599.63 702.59 34.03 4.29
Summary statistics:
credit default swap Notes: Credit default swap (CDS) categories vs agency credit ratings (ACR) categories; February 28,
spreadsa 2003, to December 31, 2010; a credit default swap spreads are in basis points

and ACR-based signals. For the best credit quality, the median swap spreads are very
close. In the quintiles, CDS-5c ⫽ 1 is 23.34 bp and ACR-5c ⫽ 1 is 30.00 bp; for the deciles,
CDS-10c ⫽ 1 is 20.44 bp and ACR-10c ⫽ 1 is 30.45 bp. For the worst credit risk quality,
the median spread level in the quintiles, ACR-5c is 1.5 times that of the CDS-5c; in the
deciles, ACR-10c is 1.6 times that of the CDS-10c. The variability of spreads increases
noticeably toward the lower credit quality in both CDS-5c/CDS-10c and ACR-5c/
ACR-10c, while this increase is more severe in agency ratings-based signals. Table IV
shows that the standard deviations of swap spreads for the best credit quality are 16.92 Credit risk
bp and 93.94 bp for CDS-5c and ACR-5c, and 15.51 bp and 81.46 bp for CDS-10c and signals
ACR-10c, respectively. For the worst credit quality, the standard deviations are 542.96
bp and 591.30 bp for CDS-5c and ACR-5c, and 685.83 bp and 702.59 bp for CDS-10c and
ACR-10c, respectively. These observations suggest that CDS market and agency
ratings credit risk assessments diverge substantially as the credit quality decreases. In
the empirical analysis section, we formally model the discrepancies in credit risk signals 205
between CDS-c and ACR-c.
Figures 1 and 2 display the evolution of median spread levels during our sample
period for each of the CDS-5c and ACR-5c, respectively. We observe that the cyclicality
of median CDS premia varies across the credit cycle for each CDS and ACR category.
To investigate the similarity or divergence of credit risk signals in the swap market
and agency ratings, we create two variables for both the quintile and the decile category
versions of the variable. The first one is a binary variable that measures the agreement
between the credit risk signals and is given by:

AGREE ⫽ 兵1 if CDS-c ⫽ ACR-c

0 otherwise
(2)

Figure 1.
Credit default swap
categories (CDS-5c)
JRF
17,2

206

Figure 2.
Agency credit rating
categories (ACR-5c)

where CDS-c and ACR-c represent the CDS and agency credit rating categories,
respectively. We are interested in identifying the degree of agreement between the two
sources in terms of their assessment of the reference entity’s credit quality.
In the cases when credit risk signals in the swap market and agency ratings are not
the same, we want to measure how different the signals are from each other. Our second
variable measures the distance between the credit risk signals and is given by:

DIST ⫽ CDS-c ⫺ ACR-c (3)

An example of a DIST calculation is CDS-c ⫽ 5 minus ACR-c ⫽ 1 equals DIST of ⫹4.

When DIST is positive, the credit risk signal in the swap market is the worst (i.e. highest
spreads), and in the agency ratings is the best (i.e. top alphanumeric scale rating like
AAA). In this case, the level of disagreement between the two sources of credit signals is
the highest and we would like to investigate what factors might be behind it. We may
refer to such a case as the CDS spread being “expensive” relative to agencies’ assessment
of credit quality, the inference being that the cost of insurance against the credit risk of
the reference entity is high. When DIST is negative (e.g. CDS-c ⫽ 1 minus ACR-c ⫽ 5
equals DIST ⫽ ⫺4), the credit risk signal in the swap market is the best (i.e. lowest
spreads), and in the agency ratings is the worst (i.e. bottom alphanumeric scale rating
like CCC). This suggests that the CDS spread is “cheap” relative to agencies’ assessment
of credit quality, the inference being that this suggests modestly priced insurance for an
investor short the credit. In other words, DIST is the number of categories for which
CDS-c signals higher credit risk if positive and ACR-c signals lower credit risk if Credit risk
negative[12]. signals
As agency credit ratings are typically TTC credit risk assessments meant to be
insensitive to spurious or short-term fluctuations in economic variables not impacting
firm fundamentals, it is not surprising that CDS spreads react more quickly to both new
information and noise than credit ratings. Thus, the level of agreement between credit
ratings and CDS spreads should be inversely related to the amount of time that has 207
passed since the last rating adjustment. Therefore, to test this conjecture, we include in
our models a variable called time since rating change (RTCH), measuring the number of
days since the last rating adjustment by the S&P[13].

3. Empirical analysis
We model the degree of similarity or divergence between the credit risk signals from
CDS-c and ACR-c, which represent both 5-category and 10-category versions. Similarity
is modeled by:

AGREE ⫽ ␤0 ⫹ ␤1 · SP500 ⫹ ␤2 · VIX ⫹ ␤3 · TP ⫹ ␤4 · SYP ⫹ ␤5 · SDI

⫹ ␤6 · POIV ⫹ ␤7 · MCAP ⫹ ␤8 · LEV ⫹ ␤9 · EPSE
(4)
⫹ ␤10 · CRIS ⫹ ␤11 · POST ⫹ ␤12 · RTCH ⫹ ␧

where AGREE indicates the agreement between the credit risk signals in CDS categories
and in ACR categories, as in equation (2). Divergence is modeled by:

DIST ⫽ ␤0 ⫹ ␤1 · SP500 ⫹ ␤2 · VIX ⫹ ␤3 · TP ⫹ ␤4 · SYP ⫹ ␤5 · SDI

⫹ ␤6 · POIV ⫹ ␤7 · MCAP ⫹ ␤8 · LEV ⫹ ␤9 · EPSE
(5)
⫹ ␤10 · CRIS ⫹ ␤11 · POST ⫹ ␤12 · RTCH ⫹ ␩

where DIST measures the level of disagreement between the credit risk signals, i.e.
measures the cardinal difference between the CDS-c and ACR-c, as in equation (3).
Additionally, SP500 is the S&P 500 Index; VIX is the CBOE volatility index; SYP is the
swap yield premium; TP is the term premium; SDI represents the seniority of the debt;
POIV is the put option-implied volatility; MCAP is the market capitalization; LEV is the
leverage; EPSE is the earnings per share estimate; CRIS and POST are dummy
variables identifying the financial crisis and the post crisis periods, respectively; and
RTCH represents the time since rating change, as described in Section 3[14].
We estimate the models in equations (4) and (5) using multinomial logit with fixed
effects, and the maintained hypothesis is that the respective error terms ␧ and ␩ satisfy
the requisite statistical assumptions necessary for the validity of the econometric
models (4) and (5)[15]. Table V presents the estimation results[16]. We observe that, in
our models for AGREE and DIST, the coefficient estimates are all statistically
significant, signs are all economically intuitive and partial effects are generally all
economically meaningful.
Considering the coefficient estimates presented in Table V, a rising stock market
(SP500) increases the estimated chances of agreement and decreases the distance
between the CDS-c and ACR-c. We interpret this as reflective of an increased flow of
information during down markets which could augment the differences between agency
 JRF
17,2

208

Table V.

(ACR) categories
and divergence in
credit risk signals.

(CDS) categories vs
Credit default swap
Modeling similarity

agency credit ratings

CDS-5c vs ACR-5c CDS-10c vs ACR-10c
Variables AGREE DIST AGREE DIST

S&P 500 Index (SP500) 0.0402 (13.57)** ⫺0.0322 (⫺11.61)** 0.0513 (9.93)** ⫺0.0299 (⫺10.26)**
CBOE volatility index (VIX) ⫺0.0158 (⫺24.81)** 0.0118 (12.64)** ⫺0.0298 (⫺7.26)** 0.0175 (6.78)**
Term premium (TP) ⫺0.0242 (⫺12.37)** 0.0100 (0.95) ⫺0.0117 (⫺6.75)** 0.0093 (0.79)
Swap yield premium (SYP) 0.0091 (6.24)** ⫺0.0278 (⫺7.33)** 0.0184 (4.67)** ⫺0.0321 (⫺6.91)**
Senior debt indicator (SDI) 0.0196 (2.69)* ⫺0.0117 (⫺2.47)** 0.0030 (2.54)* ⫺0.0161 (⫺3.46)**
Put option implied volatility (POIV) ⫺0.0017 (⫺22.83)** 0.0102 (16.06)** ⫺0.0074 (⫺12.65)** 0.0141 (12.35)**
Market capitalization (MCAP) 0.0407 (13.97)** ⫺0.0920 (⫺22.87)** 0.0232 (9.57)** ⫺0.1224 (⫺14.77)**
Leverage (LEV) ⫺0.0120 (⫺2.85)** 0.0036 (10.17)** ⫺0.0175 (⫺2.53)* 0.0618 (9.43)**
EPS estimate (EPSE) 0.0028 (7.94)** ⫺0.0030 (⫺11.87)** 0.0084 (4.69)** ⫺0.0024 (⫺8.16)**
Crisis dummy (CRIS) ⫺0.0291 (⫺11.1)** 0.0454 (6.54)** ⫺0.0642 (⫺3.42)** 0.0835 (6.68)**
Post crisis dummy (POST) 0.0124 (2.37)* ⫺0.0103 (⫺8.23)** 0.0143 (4.33)** ⫺0.0164 (⫺7.21)**
Time since rating change (RTCH) ⫺0.0009 (⫺2.81)* 0.0010 (2.63)* ⫺0.0006 (⫺2.93)* 0.0012 (2.22)*
Constant ⫺0.2929 (⫺2.47)* 1.9174 (2.21)* ⫺0.3749 (⫺2.85)* 1.7037 (2.28)*
Number of observations 379,625 379,625 379,625 379,625
R2 between 0.6715 0.5886 0.4487 0.3924
R2 overall 0.5978 0.4906 0.3561 0.3187
Wald ␹2 576.5 350.2 247.7 183.9
Prob ⬎ ␹2 0.000 0.000 0.000 0.000

Notes: February 28, 2003, to December 31, 2010; first dependent variable is AGREE, which is a binary variable looking at the agreement between the credit risk
signals and is given by equation (2):

1 if CDS-c ⫽ ACR-c
AGREE ⫽ 兵 0 otherwise

where CDS-c and ACR-c represent the credit default swap and agency credit rating categories, respectively; we are interested in identifying the degree of agreement
between the two sources in terms of their assessment of the reference entity’s credit quality; the second dependent variable is DIST, which measures the distance
between the credit risk signals and is given by equation (3): DIST ⫽ CDS-c–ACR-c; when DIST is positive (e.g. CDS-c ⫽ 5, ACR-c ⫽ 1, and DIST ⫽ ⫹4), the credit
risk signal in the swap market is the worst (i.e., highest spreads) and in the agency ratings is the best possible (i.e., top alphanumeric scale, rating like AAA),
respectively; we estimate the models in equations (4) and (5), for AGREE and DIST, respectively, using multinomial logit with fixed effects, and the maintained
hypothesis is that the respective error terms ⑀ and ␩ satisfy the requisite statistical assumptions necessary for the validity of the econometric models (4) and
(5); ** and * denote statistical significance at the 1 and 5% levels, respectively; the z-scores are presented within parentheses
and market-based assessments of credit risk – either the agencies are maintaining a Credit risk
TTC assessment of credit risk or are reluctant to “keep up”. Alternatively, during signals
declining markets, factors that tend to drive this wedge, such as liquidity or investor risk
aversion, are more prominent and are more likely to be reflected to a greater degree in the
CDS market– based versus the agency-based assessment of credit risk. The magnitudes
of the coefficient estimates suggest that comparing the models, an increase in the S&P
500 Index increases the estimated probability of agreement and decreases the 209
probability of a one-notch distance between CDS-c and ACR-c.
We observe that an increase in the VIX (the so-called fear index) is associated with
either lower odds of agreement or greater odds of a more severe credit assessment in the
CDS market as compared to the rating agencies. It is possible that in more volatile
environments, investor risk aversion is heightened, and therefore not only is it
more likely that the CDS market does not agree with the agencies due to a more severe
measure of credit risk, but also that the “tail-risk” component of credit risk is reflected in
the CDS-c and not in the ACR-c. Magnitudes of coefficient estimates indicate that in the
CDS-5c versus the ACR-5c model, for every one point increase in the VIX, either the
probability of agreement is expected to decline by roughly 1.6 per cent, or the probability
of an additional one-notch discrepancy between CDS-5c and ACR-5c is expected to
increase by about 1.2 per cent, and in the case of the CDS-10c versus ACR-10c model, the
corresponding increase is 3 per cent and decrease is 1.8 per cent.
An increasing term premium, TP, is associated with a lower predicted likelihood of
agreement between CDS-c and ACR-c. Interestingly, the coefficient estimate of TP in the
DIST model is not statistically significant, while it is highly significant in the AGREE
model. We hypothesize that as the yield curve becomes more upwardly sloping,
investors could be extracting a higher premium for investing at the long end of the curve
(i.e. a greater time preference for immediacy). This would be directly reflected in richer
CDS pricing and higher odds that either CDS-c and ACR-c do not agree or that CDS-c are
signaling more credit risk. It is possible that rating agencies would be looking ahead
through the cycle to better economic conditions, while the CDS market would tend to
maintain a shorter-term view, which would lead to a greater probability that there is a
disagreement between CDS-c versus ACR-c. However, TP is not able to differentiate the
exact distance between signals. The size of the coefficient estimate suggests in the
CDS-5c model, substantial economic significance of the TP in explaining disagreement
in absolute terms relative to the other covariates, as a 1 per cent widening in TP implies
about a 2.4 per cent decline in the chances of agreement, and in the case of the ACR-10c
model, the corresponding decrease is 1.8 per cent.
The results for the swap yield premium, SYP¸ show that as the yield difference
between the plain interest rate swaps and the Treasuries increases, AGREE is
increasing and DIST is decreasing. Based on our variable construction, increased values
of CDS-c and ACR-c imply higher credit risk. Prior research found that increases in SYP
cause CDS spreads to increase. Therefore, we hypothesize that increased credit risk
among the inter-bank market participants is indicative of a worse credit market
environment and riskier issues; as a result, one possibility is that the rating agencies
behave differently in such environments, and are in greater harmony with the CDS
market (i.e. agency signals are more likely to agree with CDS signals). The results are
robust in terms of economic impact, in the CDS-5c versus ACR-5c model, with a 1 per
cent rise in the SYP associated with about a 0.9 per cent greater probability of agreement
JRF and approximately 1.8 per cent decreased chances of the CDS market-based credit risk
17,2 category moving one additional notch worse than the agency-based one. In the case of
the CDS-10c versus ACR-10c model, the corresponding increase is 1.8 per cent and
decrease is 3.2 per cent.
More senior issues are associated with either higher probabilities of agreement, or a
shorter distance, between the CDS-c and ACR-c. This result is readily explainable, in that less
210 senior debt is likely to be evaluated with greater skepticism among risk-averse investors in
the CDS market than analysts at rating agencies. The size of the estimates in the CDS-5c
versus ACR-5c model implies that senior issues have a 2 per cent lower probability of a
divergence in the credit risk signal between the agencies and the CDS market, as well as a 1.2
per cent higher probability of the CDS market-based credit risk category exceeding the
ACR-based one by an additional category, and in the case of the CDS-10c versus ACR-10c
model, the corresponding decrease is 0.8 per cent and increase is 1.6 per cent.
We observe in Table V that higher levels of put option-implied volatility (POIV) of the
issuer’s equity decreases the estimated probability of an agreement, and increases the
odds of a positive increment in distance, between the CDS-c and ACR-c. The intuition is
that to the extent that POIV is associated with increasing distress of the issuer, we would
expect this to be impounded to a greater extent into the CDS market’s assessment of
credit risk, as it accounts for factors like investor risk aversion and liquidity that the
rating agencies may not incorporate to the same extent into their rating assessments.
This variable is economically significant. In the CDS-5c versus ACR-5c model, a 1
percentage increase in POIV gives rise to a 0.17 per cent decline in the chances that the
CDS market and ACR agree upon the relative credit quality of the issue, as well as a 1.0
per cent rise in the estimated probability that the CDS market-based credit risk category
becomes one notch worse than the ACR-based one. In the case of the CDS-10c versus
ACR-10c model, the corresponding increase is 0.74 per cent and decrease is 1.4 per cent.
Our results suggest that the larger issuers of debt (i.e. with higher MCAP) are more
likely to have ACR-based credit risk assessments that are in harmony with the CDS
market– based ones, as well as that the latter tend to be closer to the former according to
our DIST measure. We hypothesize that for such larger firms, better quality and volume
of information mitigates factors that likely drive a wedge between the credit risk
assessments of the CDS market and that of the rating agencies – for example, credit risk
premia or liquidity effects. As an alternative and complementary explanation of the
higher observed agreement for these firms, rating agencies more intensely monitor
larger companies, and may hence adjust these companies’ ratings more frequently.
Furthermore, this effect is robust in magnitude, implying in the CDS-5c versus ACR-5c
model that roughly for every doubling in market capitalization of the issuer, it is 4.1 per
cent more probable that ACR-c agrees with the CDS-c, and 9.2 per cent more likely that
there is a reduction in the distance between the respective categories of credit quality. In
the case of the CDS-10c versus ACR-10c model, the corresponding increase is 2.3 per
cent and decrease is 12.2 per cent.
We find that more levered issuers are less likely to have debt whose credit risk is
assessed similarly by CDS market and rating agencies, and it is more likely that the
divergence between CDS-c and ACR-c will widen. This supports the hypothesis that
greater leverage is associated with an increase in the default boundary and therefore an
elevated risk of default, which in turn increases risk premia demanded in the CDS
market above and beyond risk of credit loss that is the focus of the ratings agencies. Our
estimates are economically meaningful, as we see in the CDS-5c versus ACR-5c model Credit risk
that every 1 per cent increase in leverage (LEV) results in a 1.2 per cent reduction in the signals
likelihood that the CDS market and ACR agree upon the relative credit quality, as well as
an 0.4 per cent increase in the estimated probability that the discrepancy between CDS-c
and ACR-c widens by one notch in the direction of a worse CDS-based credit risk signal
relative to an agency-based one. In the case of the CDS-10c versus ACR-10c model, the
corresponding increase is 1.8 per cent and decrease is 6.2 per cent. 211
An improving earnings per share estimate (EPSE) among analysts is associated with a
greater probability of agreement, as well as a lesser probability of a one-notch difference,
between CDS-c and ACR-c. We interpret this result as correlated to analysts’ better EPS
estimates of the issuer: the implied diminished credit risk leads to a decline in the factors that
would drive a wedge between the CDS-c and the ACR-c, such as investor risk aversion. The
coefficient estimates indicate that in the CDS-5c versus ACR-5c model, every 1 per cent
increase in EPSE results in a 0.3 per cent greater estimated probability of agreement, or a 0.3
per cent lesser chance of a one-notch disagreement, in the CDS-market and ACR-based credit
risk assessments. In the case of the CDS-10c versus ACR-10c model, the corresponding
increase is 0.8 per cent and decrease is 0.2 per cent.
We observe that during crisis, there has been a decrease in the probability of
agreement and an increase in the discrepancy between the credit risk signals in the CDS
market and agency ratings. This makes sense: during financial crisis and economic
downturn periods, issuers are generally in a state of worse credit quality, wherein
factors such as investor risk aversion or liquidity give rise to more severe credit
assessments in the CDS market vis-à-vis the rating agencies. The coefficient estimates
suggest that in the CDS-5c versus ACR-5c model, a 10 per cent increase in the financial
crisis period experienced about a 1.2 per cent decrease in the estimated probability of
agreement and a 4.5 per cent increase in the probability of a one-notch distance between
CDS-c and ACR-c. In the case of the CDS-10c versus ACR-10c model, the corresponding
decrease is 6.4 per cent and increase is 8.4 per cent.
Finally, considering the RTCH (time since rating change) variable, our empirical
results show that the level of agreement decreases and distance between the credit risk
categories increases as more time has passed since the last rating adjustment. That is,
the level of agreement between credit ratings and CDS spreads is inversely related to the
amount of time that has passed since the last rating adjustment. The coefficient
estimates suggest that in the CDS-5c versus ACR-5c model, an additional quarter since
a rating change equates to about a 6 per cent decrease in the estimated probability of
agreement and a 6 per cent increase in the probability of a one-notch distance between
CDS-c and ACR-c. In the case of the CDS-10c versus ACR-10c model, the corresponding
decrease is 8 per cent and increase is 9 per cent.
To verify the robustness of our results regarding the relationship between credit risk
signals from the two sources in our analysis, we directly regress CDS-c on ACR-c, as well
as on the other factors from their respective models (4) and (5) as independent variables
using the following model[17]:

CDS-c ⫽ ␤0 ⫹ ␤1 · ACR-c ⫹ ␤2 · SP500 ⫹ ␤3 · VIX ⫹ ␤4 · TP

⫹ ␤5 · SYP ⫹ ␤6 · SDI ⫹ ␤7 · POIV ⫹ ␤8 · MCAP ⫹ ␤9 · LEV (6)
⫹ ␤10 · EPSE ⫹ ␤11 · CRIS ⫹ ␤12 · POST ⫹ ␤13 · RTCH ⫹ ␰
JRF where CDS-c and ACR-c represent the CDS and agency credit rating categories,
17,2 respectively. The first column in Table VI presents the estimated coefficients in
equation (6) using the five-category classifications, CDS-5c based on quintiles of spread
levels and ACR-5c based on the S&P-assigned long-term ratings rescaled into five
categories. The second column in Table VI presents the results when we use the
10-category classifications, CDS-10c based on deciles of spread levels and ACR-10c
212 based on the deciles of distribution of agency ratings in our data set. Other variables in
equation (6) are identical to those defined in our equations (4) and (5).
We observe that our empirical results hold and that the coefficient estimate of
ACR-5c on CDS-5c is 0.8735 (ACR-10c on CDS-10c is 0.7912), both highly statistically
significant, indicating a very high degree of correspondence between the credit signals
from two sources. Furthermore, not only are the signs of the parameter estimates on the
explanatory variables economically intuitive and statistically significant, they are of
similar magnitudes (and in some cases very close) across CDS-5c and CDS-10c models.

CDS category
Variablesa (CDS-5c)b (CDS-10c)c

Agency rating category (ACR-5c) 0.8735 (16.55)**

Agency rating category (ACR-10c) 0.7912 (8.49)**
S&P 500 index (SP500) ⫺0.2246 (⫺9.17)** ⫺0.1915 (⫺8.72)**
CBOE volatility index (VIX) 0.0641 (11.25)** 0.0756 (12.82)**
Term premium (TP) 0.0830 (3.56)** 0.0848 (2.94)*
Swap yield premium (SYP) 0.0638 (5.82)** 0.0552 (4.85)**
Senior debt indicator (SDI) ⫺0.0118 (⫺7.47)** ⫺0.0158 (⫺6.13)**
Put option implied volatility (POIV) 0.0092 (12.77)** 0.0134 (14.25)**
Market capitalization (MCAP) ⫺0.0871 (⫺4.74)** ⫺0.0713 (⫺3.62)**
Leverage (LEV) 0.0138 (8.68)** 0.0167 (7.56)**
EPS estimate (EPSE) ⫺0.0032 (⫺3.85)** ⫺0.0037 (⫺4.29)**
Crisis dummy (CRIS) 0.3720 (5.53)** 0.3748 (5.91)**
Post crisis dummy (POST) ⫺0.8232 (⫺6.96)* ⫺0.8336 (⫺6.36)**
Time since rating change (RTCH) ⫺0.0001 (⫺2.34)* ⫺0.0001 (⫺2.96)*
Constant 0.6670 (0.45) 0.4408 (0.32)
Number of observations 379,625 379,625
R2 between 0.5973 0.4365
R2 overall 0.4850 0.3496
Wald ␹2 306.7 245.2
Prob ⬎ ␹2 0.000 0.000

Table VI. Notes: February 28, 2003, to December 31, 2010; in this table, we model the relationship between credit
Modeling the risk signals from the two sources in our analysis by directly regressing CDS spread categories (CDS-c)
relationship between on credit rating categories (ACR-c), as well as on the other firm-specific and market
credit risk signals: variables; a variables from equation (6) are similarly defined in equations (4) and (5); b 5-category
CDS categories as a classifications (i.e., CDS-5c), which are based on quintiles of spread levels, and ACR-5c, which are based
function of ACR on the S&P assigned long-term ratings rescaled into five categories; c 10-category classifications (i.e.,
categories as well as CDS-10c), which are based on deciles of spread levels, and ACR-10c, which are based on the deciles of
firm-specific and distribution of agency ratings in our data set; ** and * denote statistical significance at the 1 and 5%
market variables levels, respectively; the z-scores are presented within parentheses
Our empirical tests confirm our hypothesis that the differences between agency credit Credit risk
ratings and credit market– based assessments can be explained by variables in line with signals
accepted economic theory. We have built two alternative models of this differential and
have shown that both challengers have the same qualitative features in terms of the
statistical and economic significance of this set of variables. Then we have directly
modeled our two measures (the agreement indicator and the notch difference) and this
robustness exercise has supported our results. In the process, we have provided a 213
potential tool for practitioners who wish to bridge the divide between PIT and TTC
credit risk measures, but who may be dealing with non-tradable instruments.

4. Conclusion
Before the inception of CDS trading, ratings assigned by agencies were the only signal
of credit risk of fixed-income instruments. In that credit ratings rank order the relative
credit risk of an entity, these assessments should be closely related to the CDS spreads
on the corresponding defaultable instruments with respect to the obligors.
However, these measures differ in a fundamental sense: CDS spreads are a real-time
market signal regarding immediate creditworthiness, whereas credit ratings are a
discrete assessment of credit risk over a longer horizon. Callen et al. (2009) observe that
while credit ratings may in fact have a close relation to CDS spreads, nevertheless with
respect to obligors sharing a common credit rating, we observe much variation in the
latter. In this vein, Cizel (2013) argues that the CDS spreads represent a more
market-based measure of a firm’s credit risk relative to credit risk ratings. If, on the one
hand, CDS spreads represent an element of pure credit risk (i.e. the danger of obligor
degradation in credit quality or default), and if, on the other hand, credit ratings are a
relative default risk metric, then there should be a correlation between the market price
of credit risk and the credit rating assigned an obligor. Herein our research agenda has
the objective to build models of the link between CDS spreads and credit ratings,
through defining alternative agreement measures between these, as well as identifying
economically intuitive covariates that explain such differences.
Our model can be used by investors in debt instruments which do not have CDS
written on them or which have illiquid CDS contracts. The findings replicate
market-based, PIT credit risk signals based on the factors (both market-based and
firm-specific) in our model, and can be used to augment TTC credit risk assessments[18].
We analyze the differences in the relative credit risk assessment from CDS-based and
agency ratings– based measures, in part motivated by Hilscher and Wilson (2013). We
present evidence that the divergence between the credit risk signals from CDS spreads
and credit ratings can be explained by factors which the rating agencies may consider
differently than credit market participants.
We provide a comprehensive analysis of the differences in the relative credit risk
assessment as implied by the ACR-c versus the CDS-c, using two models to better
explain the ACR-c being equal to the CDS-c, or the CDS-c being a more severe
assessment of credit risk than the ACR-c. The results are statistically and economically
significant, and had signs on (magnitudes of) coefficient estimates that are economically
intuitive (significant). The VIX, term premium or the put option-implied volatility of the
associated equity decrease the chances of agreement and increase the relative
pessimism on the part of the CDS market. Conversely, the level of the S&P 500 index,
market capitalization or the EPS estimate of the reference entity, as well as the seniority
JRF of the debt, are always negatively related to agreement and disagreement of CDS-market
17,2 and agency ratings credit risk signals.
Therefore, our analysis suggests that for investors and other stakeholders in the
credit markets having a more PIT rather than TTC orientation, modifications need to be
made to the method of determining relative credit risk that incorporate market-based
signals such as the information contained in CDS spreads. Moreover, if the credit in
214 question is not traded in the CDS market (or is an illiquid debt issue), there are factors
available to augment TTC measures like agency rating assessments[19] that will mimic
the information in the CDS-based signals. Said differently, these results suggest that
agency credit ratings of relative riskiness of a reference entity do not always correspond
with assessments by CDS spreads, as the price of risk is a function of additional macro
and micro factors that can be explained using statistical analysis. Finally, the models we
developed and their empirical findings could be used to analyze issues surrounding the
pricing of CDSs and examine the policies of credit rating agencies[20]. For example, in
cases of a rated pool of credits or a structured vehicle rated by an agency which has a
market offering price but is not actively traded, our model could be used as a benchmark
to provide an alternative rating that could be used to judge if the agencies are pricing
aggressively or not with respect to the market.

Notes
1. The three most prominent NRSROs are S&P, Moody’s Investors Services and Fitch.
2. Ratings designed to reflect borrowers’ fundamental creditworthiness are sometimes termed
the TTC philosophy of ratings, as compared to the PIT orientation of market-based rating
schemes. The TTC ratings are generally favored by risk managers and prudential
supervisors for capital management purposes because they result in less cyclicality in capital
measures. However, as PIT ratings give more reactive signals, these parties sometimes prefer
PIT ratings for account management purposes (Resti and Sironi, 2007).
3. Hull et al. (2004) use Moody’s ratings and separate the data into three categories: Aaa-Aa, A
and Baa; however, they do not consider high-yield names.
4. Hull et al. (2004) argue the best risk-free curve to use when pricing CDSs lies somewhere
between the Treasury and the swap curve. However, many practitioners prefer to use the
swap curve, as this is the curve used most often in derivative pricing. Fabozzi et al. (2007)
examine the effects of various variables on pricing the CDS. They find that in addition to
risk-free interest rate and ratings, the liquidity factors especially affect the CDS spreads.
5. Cizel (2013) suggests that these instruments represent the most liquid and traded category of
CDS contracts. .
6. One of the reasons for choosing S&P ratings as compared to alternatives such as Moody’s
ratings is that they have certain advantages: Güttler (2011) presents evidence that S&P
assigns ratings in a timelier manner than Moody’s, as the tendency toward rating
convergence is found to be stronger for S&P than for Moody’s. Among the three rating
agencies of S&P, Moody’s and Fitch, Cizel (2013) finds statistically and economically
significant responses only to S&P announcements.
7. The estimate of the earnings per share is the average of Bloomberg’s collection of analysts’
EPS forecasts.
 8. Construction of our crisis and post-crisis dummy variables are based on Huang et al. (2012). Credit risk
Their analysis suggests that by mid-2009, market conditions had returned to normalcy.
signals
9. Norden and Weber (2004) in their study of the effect of agency “watch list” entries on credit
risk assessments use Bloomberg rating changes in a similar scale. As credit ratings change
overtime, the number of names in each category also changes over time.
10. Huang et al. (2012) use two groups: investment-grade AAA to BBB– and noninvestment-grade
BB⫹ and below. Cizel (2013) uses five broad rating categories: AAA/AA (Aaa, Aa), A, BBB (Baa), 215
BB (Ba) and B or below. In an alternative approach, Castellano and Giacometti (2012) calibrate an
implied rating (IR) model on CDS spreads that aims to determine the optimal IR bounds,
minimizing the average distance between each spread and the crossed boundary, for all CDSs and
all rating classes.
11. Di Cesare (2006) examines the ability of market-based indicators to anticipate the rating-related
decisions of agencies and finds that CDS spreads predict negative rating events.
12. Hornik et al. (2010) study all three aspects of proximity – association, agreement and bias – in
the context of rating systems, concluding that all of these are of relevance. Furthermore, they
argue that although these aspects are not necessarily independent, it is evident that none of
the three aspects is redundant.
13. We thank an anonymous referee for recommending this variable in our analysis.
14. In our empirical analysis, we use the natural log of firm-specific variables and market
indicators.
15. This method is similar to the calibration modeling in Das and Hanouna (2008).
16. Our estimation methodology uses heteroscedasticity and autocorrelation consistent (HAC) robust
standard errors, which are derived from either bootstrap or jackknife methods. We use customary
post-estimation tests to verify that our model specifications do not suffer from autocorrelation,
non-stationarity, heteroscedasticity and multicollinearity. Results are available upon request.
17. We thank an anonymous referee for suggesting this model for robustness checks.
18. For example, consider bank evaluating the sale or securitization of middle market obligors,
which do not have CDS quotes for their debt, but the bank has its internal ratings for those
that map to the agency credit ratings. In such a case, our calibrated model could be used to
produce a proxy CDS market-based credit signal that would be used as a benchmark or
baseline for pricing this deal.
19. Note that this could also apply to bank loans having internal TTC ratings, where the CDS is
illiquid or nonexistent, but the some of the variables in our model are available.
20. We thank an anonymous referee for pointing out such potential uses of our models and empirical
results.

References
Basel Committee on Banking Supervision (BCBS) (2006), International Convergence of Capital
Measurement and Capital Standards: A Revised Framework, Comprehensive Version,
Bank for International Settlements, Basel.
Basel Committee on Banking Supervision (BCBS) (2010), Basel III: A Global Regulatory Framework for
More Resilient Banks and Banking Systems, Bank for International Settlements, Basel.
Benzschawel, T. and Corlu, A. (2011), “Credit default swaps: a cash flow analysis”, Journal of Fixed
Income, Vol. 20 No. 3, pp. 40-55.
JRF Callen, J.L., Livnat, J. and Segal, D. (2009), “The impact of earnings on the pricing of credit default
swaps”, The Accounting Review, Vol. 84 No. 5, pp. 1363-1394.
17,2
Cao, C., Fan, Y. and Zhaodong, Z. (2010), “The information content of option-implied volatility for
credit default swap valuation”, Journal of Financial Markets, Vol. 13 No. 3, pp. 321-343.
Carr, P. and Wu, L. (2010), “Stock options and credit default swaps: a joint framework for valuation
and estimation”, Journal of Financial Econometrics, Vol. 8 No. 4, pp. 409-449.
216 Castellano, R. and Giacometti, R. (2012), “Credit default swaps: implied ratings versus official
ones”, 4OR: Quarterly Journal of Operations Research, Vol. 10 No. 2, pp. 163-180.
Childs, M. (2014, January 30), “The incredible shrinking credit-default swap market”, Bloomberg
Business, available at: www.bloomberg.com/bw/articles/2014-01-30/credit-default-swap-
market-shrinks-by-half (accessed 13 November 2015).
Cizel, J. (2013), “Are credit rating announcements contagious? Evidence on the transmission of
information across industries in credit default swap markets”, Journal of Fixed Income,
Vol. 23 No. 2, pp. 27-60.
Cox, J.C., Ingersoll, J.E. and Ross, S.A. (1981), “A Re-examination of traditional hypotheses about
the term structure of interest rates”, Journal of Finance, Vol. 36 No. 4, pp. 769-799.
Daniels, K.N. and Shin-Jensen, M. (2005), “The effect of credit ratings on credit default swap
spreads and credit spreads”, Journal of Fixed Income, Vol. 15 No. 3, pp. 16-33.
Das, S.R. and Hanouna, P. (2008), “Implied recovery”, available at: http://ssrn.com/abstract⫽102
8612 (accessed 13 November 2015).
Das, S.R. and Hanouna, P. (2009), “Hedging credit: equity liquidity matters”, Journal of Financial
Intermediation, Vol. 18 No. 1, pp. 112-123.
Di Cesare, A. (2006), “Do market-based indicators anticipate rating agencies? Evidence for
international banks”, Economic Notes, Vol. 35 No. 1, pp. 121-150.
Dodd-Frank Wall Street Reform and Consumer Protection Act (2010), Public Law 111-203, 111th
Congress, July 21, 2010, 124 Stat. 1376.
Eaglesham, J., Neumann, J. and Perez, E. (2013), “US sues S&P over ratings”, The Wall Street
Journal, 5 February, available at: www.wsj.com/articles/SB10001424127887324445904578
284064003795142 (accessed 13 November 2015).
Fabozzi, F.J., Cheng, X. and Chen, R. (2007), “Exploring the components of credit risk in credit
default swaps”, Finance Research Letters, Vol. 4 No. 1, pp. 10-18.
Güttler, A. (2011), “Lead-lag relationships and rating convergence among credit rating agencies”,
Journal of Credit Risk, Vol. 7 No. 1, pp. 95-119.
Hilscher, J. and Wilson, M.I. (2013), “Credit ratings and credit risk: is one measure enough?”
Working Paper AFA 2013 San Diego Meetings, available at: http://ssrn.com/abstract⫽14
74863 (accessed 13 November 2015).
Hornik, K., Jankowitsch, R., Leitner, C., Lingo, M., Pichler, S. and Winkler, G. (2010), “A latent variable
approach to validate credit rating systems”, in Rösch, D. and Scheule, H. (Eds), Model Risk:
Identification, Measurement and Management, Risk Books, London, pp. 277-296.
Huang, A.Y., Shen, C.-H. and Chen, C.C. (2012), “The impact of major events from the recent
financial crisis on credit default swaps”, Journal of Fixed Income, Vol. 21 No. 3, pp. 31-43.
Hull, J., Predescu, M. and White, A. (2004), “The relationship between credit default swap spreads,
bond yields, and credit rating announcements”, Journal of Banking and Finance, Vol. 28
No. 11, pp. 2789-2811.
Hull, J.C. (2009), “The credit crunch of 2007: what went wrong? Why? What lessons can be
learned?”, Journal of Credit Risk, Vol. 5 No. 2, pp. 3-18.
International Organization of Securities Commissions (IOSCO) (2004), Code of Conduct Credit risk
Fundamentals for Credit Rating Agencies, available at: www.iosco.org/library/pubdocs/
pdf/IOSCOPD180.pdf (accessed 13 November 2015). signals
International Swaps and Derivatives Association (ISDA) (2010), 2010 Mid-year Market Survey,
ISDA Market Surveys 1995-2010, available at: www2.isda.org/functional-areas/research/
surveys/market-surveys/ (accessed 13 November 2015).
Jacobs, M., Jr and Karagozoglu, A.K. (2011), “Modeling ultimate loss given default on corporate 217
debt”, Journal of Fixed Income, Vol. 21 No. 1, pp. 6-20.
Liu, J., Longstaff, F.A. and Mandell, R.E. (2006), “The Market price of risk in interest rate swaps:
the roles of default and liquidity risks”, Journal of Business, Vol. 79 No. 5, pp. 2337-2359.
Mahfoudhi, R. (2011), “Credit default swap trees”, Journal of Credit Risk, Vol. 7 No. 3, pp. 3-37.
Merton, R. (1974), “On the pricing of corporate debt: the risk structure of interest rates”, Journal of
Finance, Vol. 29 No. 2, pp. 449-470.
Micu, M., Remolona, E. and Wooldridge, P. (2006), “The price impact of rating announcements:
which announcements matter?”, working paper 207, Bank for International Settlements,
Basel, June, available at: www.bis.org/publ/work207.pdf (accessed 13 November 2015).
Norden, L. and Weber, M. (2004), “Informational efficiency of credit default swap and stock
markets: the impact of credit rating announcements”, Journal of Banking and Finance,
Vol. 28 No. 11, pp. 2813-2843.
Resti, A. and Sironi, A. (2007), Risk Management and Shareholders’ Value in Banking: From Risk
Measurement Models to Capital Allocation Policies, John Wiley & Sons, West Sussex.
Ross, A. (2008), “JP Morgan pays $2 a share for bear stearns”, The New York Times, 17 March.
Schneider, P., Sogner, L. and Veza, T. (2007), “Jumps and recovery rates inferred from corporate
CDS premia”, Working Paper, European Financial Management Association, Norfolk, VA,
available at: www.efmaefm.org/0EFMAMEETINGS/EFMAANNUAL MEETINGS/2007-
Austria/papers/0463.pdf (accessed 13 November 2015).
Stephens, P. (2012), “Downgrade the rating agencies”, Financial Times, 19 January.
Stulz, R.M. (2010), “Credit default swaps and the credit crisis”, Journal of Economic Perspectives,
Vol. 24 No. 1, pp. 73-92.
Villouta, C. (2006), “Empirical study of liquidity effects in the relation between corporate credit
spread and credit default swaps”, available at: http://ssrn.com/abstract⫽906165 (accessed
13 November 2015).
Zhang, B.Y., Zhou, H. and Zhu, H. (2009), “Explaining credit default swap spreads with the equity
volatility and jump risks of individual firms”, Review of Financial Studies, Vol. 22 No. 12,
pp. 5099-5131.

Corresponding author
Ahmet K. Karagozoglu can be contacted at: finakk@hofstra.edu

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com